layer_type stringclasses 8
values | shard_dir stringlengths 15 32 | sample_idx int64 0 799 |
|---|---|---|
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0005 | 80 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0187 | 179 |
lobe | lobe/shard_0208 | 305 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0092 | 569 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0247 | 274 |
channel:SH_DISTAL | channel_sh_distal/shard_0143 | 339 |
delta | delta/shard_0008 | 215 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0068 | 409 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0062 | 479 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0090 | 282 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0145 | 59 |
delta | delta/shard_0203 | 118 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0214 | 186 |
delta | delta/shard_0217 | 456 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0128 | 10 |
lobe | lobe/shard_0179 | 376 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0254 | 346 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0251 | 206 |
lobe | lobe/shard_0058 | 173 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0129 | 82 |
channel:SH_DISTAL | channel_sh_distal/shard_0177 | 235 |
delta | delta/shard_0132 | 60 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0018 | 103 |
delta | delta/shard_0105 | 510 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0226 | 198 |
lobe | lobe/shard_0125 | 23 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0192 | 338 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0043 | 315 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0246 | 250 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0077 | 341 |
delta | delta/shard_0119 | 384 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0175 | 352 |
delta | delta/shard_0157 | 395 |
lobe | lobe/shard_0091 | 203 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0150 | 272 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0144 | 124 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0174 | 319 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0226 | 114 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0113 | 543 |
delta | delta/shard_0049 | 303 |
lobe | lobe/shard_0130 | 699 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0104 | 300 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0214 | 297 |
lobe | lobe/shard_0058 | 209 |
channel:SH_DISTAL | channel_sh_distal/shard_0134 | 213 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0121 | 441 |
lobe | lobe/shard_0013 | 486 |
delta | delta/shard_0177 | 271 |
delta | delta/shard_0226 | 113 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0001 | 309 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0202 | 163 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0210 | 299 |
delta | delta/shard_0092 | 269 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0117 | 232 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0010 | 573 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0158 | 161 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0065 | 353 |
lobe | lobe/shard_0069 | 483 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0254 | 265 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0144 | 357 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0004 | 146 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0082 | 76 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0168 | 338 |
lobe | lobe/shard_0068 | 554 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0144 | 354 |
delta | delta/shard_0206 | 23 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0242 | 104 |
channel:SH_DISTAL | channel_sh_distal/shard_0181 | 316 |
channel:SH_DISTAL | channel_sh_distal/shard_0252 | 239 |
delta | delta/shard_0096 | 249 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0154 | 265 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0072 | 416 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0004 | 9 |
delta | delta/shard_0007 | 350 |
lobe | lobe/shard_0127 | 703 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0060 | 379 |
channel:CB_LABYRINTH | channel_cb_labyrinth/shard_0191 | 165 |
channel:SH_PROXIMAL | channel_sh_proximal/shard_0187 | 309 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0075 | 243 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0078 | 56 |
channel:MEANDER_OXBOW | channel_meander_oxbow/shard_0022 | 130 |
channel:SH_DISTAL | channel_sh_distal/shard_0049 | 14 |
delta | delta/shard_0087 | 210 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0002 | 332 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0128 | 88 |
channel:PV_SHOESTRING | channel_pv_shoestring/shard_0104 | 264 |
delta | delta/shard_0248 | 8 |
lobe | lobe/shard_0160 | 756 |
lobe | lobe/shard_0241 | 271 |
delta | delta/shard_0143 | 326 |
channel:SH_DISTAL | channel_sh_distal/shard_0238 | 118 |
lobe | lobe/shard_0040 | 268 |
delta | delta/shard_0254 | 364 |
lobe | lobe/shard_0236 | 284 |
channel:CB_JIGSAW | channel_cb_jigsaw/shard_0204 | 20 |
delta | delta/shard_0175 | 555 |
channel:SH_DISTAL | channel_sh_distal/shard_0224 | 54 |
lobe | lobe/shard_0002 | 230 |
delta | delta/shard_0124 | 459 |
delta | delta/shard_0008 | 116 |
- Quick stats
- File structure
- Layer types
- What varies between samples
- Voxel arrays per sample
- Slim parquet (
params_slim.parquet) — the training conditioning - Full parquet (
params.parquet) — for reproducibility - Splits
- Cell-physical-size note
- Loading example (PyTorch)
- Recommended preprocessing for flow-matching training
- Geological caveats
- Citation
- License
- Contact / Issues
Siliciclastic Reservoirs
Released by SciLM.ai: https://www.scilm.ai
Engine: ResMill (pip install resmill, version 0.2.0) ·
Model: ResFlow ·
Benchmark: ResBench ·
Paper: NeurIPS 2026, Track on Evaluations and Datasets (citation)
1,000,000 synthetic 3D siliciclastic-reservoir geology cubes generated from rule-based sedimentological simulations (turbidite lobes + 6 fluvial-channel architectures + a distributary-tree delta). Each cube is a (64, 64, 32) voxel window cut at a random position out of a (128, 128, 64) simulated volume, so every cube is a piece of a larger reservoir and its faces cut through sand bodies the way the boundaries of a modelled interval do.
Designed to train conditional generative models (flow matching, diffusion, etc.) of subsurface geology under interpretable physical conditioning.
Quick stats
- 1,000,000 samples across 8 reservoir architectures
- Cube shape:
(64, 64, 32)voxels —(x, y, z),zis depth (0 = base) - 4 voxel arrays per sample: binary facies, 6-class facies, porosity, permeability
- Compact slim parquet (9–12 cols) and full reproducibility parquet (34–108 cols)
- Captions for text-conditioning experiments
- Every cube holds sand (realised net-to-gross > 0)
- Fully reproducible — every volume regenerable from
(seed, params), every window from itscrop_*columns
File structure
SiliciclasticReservoirs/
├── README.md ← this file
├── DATASHEET.md ← Datasheet for Datasets (Gebru et al. 2018)
├── splits/ ← deterministic 90/5/5 split, stratified by layer_type
│ ├── train.parquet ← 900,000 rows
│ ├── validation.parquet ← 50,000 rows
│ └── test.parquet ← 50,000 rows
├── lobe/ ← 200,000 samples, 256 shards
│ ├── shard_0000/
│ │ ├── facies.npy (n, 64, 64, 32) int8 — binary 0/1
│ │ ├── facies_alluvsim.npy (n, 64, 64, 32) int8 — 6-class -1..4
│ │ ├── poro.npy (n, 64, 64, 32) float16 — porosity in [0, 0.5]
│ │ ├── perm.npy (n, 64, 64, 32) float16 — permeability in mD [0, 60000]
│ │ ├── params_slim.parquet ← training conditioning (use this)
│ │ └── params.parquet ← full physics for reproducibility
│ ├── shard_0001/
│ └── …
├── channel_pv_shoestring/ ← 100,000 samples, 256 shards
├── channel_cb_labyrinth/ ← 100,000 samples, 256 shards
├── channel_cb_jigsaw/ ← 150,000 samples, 256 shards
├── channel_sh_distal/ ← 100,000 samples, 256 shards
├── channel_sh_proximal/ ← 100,000 samples, 256 shards
├── channel_meander_oxbow/ ← 100,000 samples, 256 shards
└── delta/ ← 150,000 samples, 256 shards
256 shards per family, each about 390 samples (100,000-sample families, ~300 MB), 585 (150,000, ~460 MB) or 780 (lobes, ~615 MB).
Layer types
layer_type |
architecture | samples |
|---|---|---|
lobe |
turbidite lobe (deep-water gravity-flow deposit) | 200,000 |
channel:PV_SHOESTRING |
paleo-valley shoestring sandstone | 100,000 |
channel:CB_JIGSAW |
channel-and-bar bodies, jigsaw connectivity | 150,000 |
channel:CB_LABYRINTH |
channel-and-bar bodies, labyrinthine connectivity | 100,000 |
channel:SH_DISTAL |
distal sheet-sandstone | 100,000 |
channel:SH_PROXIMAL |
proximal sheet-sandstone | 100,000 |
channel:MEANDER_OXBOW |
multi-storey meander-belt with oxbow plugs | 100,000 |
delta |
distributary-tree delta with mouth bars | 150,000 |
These follow Pyrcz & Deutsch's standard fluvial-architecture taxonomy ("User Guide to the Alluvsim Program", 2004).
What varies between samples
Every family is a Sobol sweep (seed 42) of its architecture's parameters. Beyond the classic knobs (net-to-gross target, sinuosity, avulsion probabilities, mud-plug fraction, levee and splay geometry, lobe size and aspect), the sweep covers:
- Channel size: depth 3 to 16 m, log-uniform, with the width-to-depth ratio sampled per family (8 to 26), so channels span 2.5 to 41 cells wide and 3 to 16 cells thick in a 32-cell-tall cube.
- Vertical stacking: the number of channel levels follows a sampled aggradation ratio (level spacing over channel depth, 0.7 to 1.4), so stacked channels range from cutting into each other to leaving up to 0.4 depth of mud between them, and the simulated column is always spanned floor to roof.
- Entry points: one to three river entries on the upstream edge, and a sampled entry scatter from a nodal point (all channels through one spot) to entries spread over the whole edge.
- Porosity texture: a correlated Gaussian texture inside every sand body on top of the upward-fining ramp, with permeability following through a Kozeny-Carman coupling.
- Delta: distributary trees with discharge split at every bifurcation (width ∝ Q^0.5, depth ∝ Q^0.4), branches that may re-merge, mouth bars at every terminus and bifurcation, a discharge-dependent prograding front, and 4 to 21 stacked generations.
Voxel arrays per sample
| file | dtype | shape | description |
|---|---|---|---|
facies.npy |
int8 |
(N, 64, 64, 32) |
binary facies: 0 = mud, 1 = sand. Primary signal for generative training. |
facies_alluvsim.npy |
int8 |
(N, 64, 64, 32) |
6-class facies preserving architectural sub-detail: -1 = FF (overbank fines), 0 = FFCH (mud plug), 1 = CS (crevasse splay), 2 = LV (levee), 3 = LA (lateral-accretion / point bar / mouth bar), 4 = CH (active channel). |
poro.npy |
float16 |
(N, 64, 64, 32) |
porosity, range [0, 0.5] |
perm.npy |
float16 |
(N, 64, 64, 32) |
permeability in millidarcies (mD), range [0, 60000]. Use log10(perm + 1e-3) for ML — perm spans 5+ decades. |
All four arrays are row-aligned within a shard (sample i of facies corresponds to row i of the parquets and to sample i of all other arrays).
Slim parquet (params_slim.parquet) — the training conditioning
This is what you train on as conditioning. It carries the geological parameters that vary across samples.
Universal columns (all 8 layer types)
| column | dtype | description |
|---|---|---|
layer_type |
str | one of the 8 strings above |
caption |
str | human-readable description, e.g. "Paleo-valley shoestring reservoir with sinuosity 1.55, channel depth 9.7 m, …". Useful for text-conditioned variants. |
ntg |
float | realised net-to-gross of the stored cube, range (0, 1] |
poro_ave |
float | realised mean porosity over the cube's sand cells |
perm_ave |
float | realised mean of log10(perm in mD) over the cube's sand cells |
azimuth |
float | regional flow direction, compass-CW degrees [0, 360) |
width_cells |
float | characteristic horizontal extent in cells. Lobe = 2 × semi-minor axis. Channel/delta = full channel width. |
depth_cells |
float | characteristic vertical extent in cells. Lobe = thickness dh_ave. Channel/delta = mCHdepth. |
Family-specific columns (NULL outside family)
| column | dtype | family | description |
|---|---|---|---|
asp |
float | lobe | ellipse aspect ratio (major/minor), [1.0, 2.5] |
mCHsinu |
float | channel + delta | sinuosity (path / straight-line), [1.05, 1.85] |
probAvulInside |
float | channel + delta | per-event in-belt avulsion probability, [0, 0.8] (recorded for the delta but not used by its tree mode) |
mFFCHprop |
float | channel + delta | abandoned-channel mud-plug fraction, [0, 0.7] |
trunk_length_fraction |
float | delta | proximal-trunk fraction, [0.1, 0.5] |
Total: 9 cols for lobe rows, 11 for channel rows, 12 for delta rows. Rows from different families share the universal columns and have NULL in non-applicable family columns.
Full parquet (params.parquet) — for reproducibility
34 columns for lobe rows, 87 for channel rows, 108 for delta rows: every sampled physics parameter, the per-event statistical knob widths (stdev*), the engine event budgets (ntime*), the realised statistics, and the window provenance:
| column | description |
|---|---|
seed |
per-sample RNG seed of the simulated volume |
crop_x0, crop_y0, crop_z0 |
origin of this cube inside its (128, 128, 64) volume; crop_nx, crop_ny, crop_nz = 64, 64, 32 |
crop_seed |
seed of the window draw (42) |
source_shard, source_row |
the rank shard and row of the volume in the generation output |
A volume is regenerated bit-for-bit from (seed, params) with the open-source engine ResMill (see its REPRODUCIBILITY.md), and the cube is volume[crop_x0:crop_x0+64, crop_y0:crop_y0+64, crop_z0:crop_z0+32]. The window origin is numpy.random.default_rng([crop_seed, seed]): x and y uniform over 0..64, z uniform over 1..31, redrawn from the same stream while the window has no sand.
Splits
splits/{train,validation,test}.parquet define a deterministic 90 / 5 / 5 partition stratified by layer_type (each split has all 8 architectures in proportion). Generated with seed 42.
Each row is (layer_type, shard_dir, sample_idx) pointing into one of the data shards.
Cell-physical-size note
Cube shape is identical (64×64×32) but the physical cell size differs by family:
| family | dx = dy | dz | physical extent |
|---|---|---|---|
lobe |
100 m | 1 m | 6.4 km × 6.4 km × 32 m |
channel:* and delta |
10 m | 1 m | 640 m × 640 m × 32 m |
For training in cell-units, this difference is hidden inside width_cells / depth_cells. The model can ignore physical scale and just train on the cube as-is.
Loading example (PyTorch)
import numpy as np
import pyarrow.parquet as pq
from pathlib import Path
import torch
from torch.utils.data import IterableDataset
class ReservoirShardDataset(IterableDataset):
"""One worker walks an assigned subset of shards, yields (cube, params)."""
def __init__(self, root: Path, split: str = "train"):
self.root = Path(root)
# Read split index
self.index = pq.read_table(self.root / f"splits/{split}.parquet").to_pylist()
def __iter__(self):
worker = torch.utils.data.get_worker_info()
my_rows = (self.index if worker is None
else self.index[worker.id::worker.num_workers])
cur_shard = None; cur_data = None
for row in my_rows:
shard_dir = self.root / row["shard_dir"]
if shard_dir != cur_shard:
cur_shard = shard_dir
cur_data = {
"facies": np.load(shard_dir / "facies.npy", mmap_mode="r"),
"poro": np.load(shard_dir / "poro.npy", mmap_mode="r"),
"perm": np.load(shard_dir / "perm.npy", mmap_mode="r"),
"params": pq.read_table(
shard_dir / "params_slim.parquet").to_pylist(),
}
i = row["sample_idx"]
yield {
"facies": np.asarray(cur_data["facies"][i], dtype=np.int8),
"poro": np.asarray(cur_data["poro"][i], dtype=np.float32),
"perm": np.asarray(cur_data["perm"][i], dtype=np.float32),
"params": cur_data["params"][i],
}
ds = ReservoirShardDataset("/path/to/SiliciclasticReservoirs", split="train")
loader = torch.utils.data.DataLoader(ds, batch_size=8, num_workers=4)
Recommended preprocessing for flow-matching training
facies→[0, 1]float: justcube.astype(np.float32)poro→ standardized:(poro - 0.15) / 0.10keeps it roughly N(0, 1)-ish for sand cellsperm→ log-perm:np.log10(np.maximum(perm, 1e-3)), then maybe standardizelayer_type→ categorical embedding (8 classes). Uselayer_type.split(":")[0]for family-level (3 classes) or full string for fine-grained (8 classes).- Other slim columns → continuous embeddings, with
null→ learned sentinel or per-family mask.
Geological caveats
- Cubes are windows of larger volumes. Sand bodies are cut by all six faces of a cube exactly as a modelled reservoir interval cuts through geology; a channel truncated at the base of a cube continues below it in the simulated volume. The window never contains the simulated floor or roof.
- Fluvial channels split but do not re-merge, except in the delta, whose distributary branches may rejoin. The braided / jigsaw look in CB_LABYRINTH / CB_JIGSAW comes from heavily overlapping channel stamps, not from explicit braid topology.
- Per-cell poro/perm follow a Walker-1992 upward-fining ramp per channel event, a per-event Kozeny-Carman-coupled draw, a per-realization "regional rock quality" multiplier, and a correlated Gaussian texture inside sand; they are not conditioned to any real reservoir.
- FACIES_PROPS base values are conservative midpoints from fluvial-reservoir literature (Pyrcz & Deutsch 2002, Allen 1965, Walker 1992), not from a specific analog.
- Net-to-gross is realised, not prescribed. Sheet and meander families reach about 0.7 to 0.9 of their sampled target in the simulated volume; the delta has no target. Condition on the realised
ntgof the cube.
See DATASHEET.md for full documentation.
Citation
Please cite the paper:
@inproceedings{baghishov2026siliciclastic,
title = {SiliciclasticReservoirs: A Million-Reservoir Dataset and Flow-Matching Foundation Model
for 3D Siliciclastic Reservoir Generation},
author = {Baghishov, Ilgar and Rustamzade, Elnara and Henkelman, Graeme and Foster, John T. and Pyrcz, Michael J.},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
year = {2026}
}
The dataset entry:
@misc{siliciclastic_reservoirs_2026,
author = {SciLM},
title = {{Siliciclastic Reservoirs}: 1M Synthetic 3D Reservoir Geology Cubes for Conditional Generative Modeling},
year = {2026},
publisher = {HuggingFace},
howpublished = {\url{https://huggingface.co/datasets/SciLM/SiliciclasticReservoirs}}
}
License
Creative Commons Attribution 4.0 International (CC-BY-4.0). Use, share, modify, redistribute — just attribute the source.
Contact / Issues
Website: https://www.scilm.ai. File issues at the dataset repository on HuggingFace.
- Downloads last month
- 5,877